StopWatch: explicit goal grounding sets SOTA on six object-goal navigation benchmarks
EugeneVinitsky · x · 2026-10-20
Researchers at Amazon FAR with NYU, UC Berkeley and CMU introduce StopWatch, an open-source navigation agent that performs explicit goal grounding before every action — solving the overlooked failure mode where policies reach goals but walk past them.
Key findings
- Diagnosis: on 3,000 HM3D-OVON valunseen episodes, an end-to-end RGB policy reaches the goal 76.3% of the time but succeeds only 52.3%; the same policy with ground-truth grounding hits 89.3%, making grounding the dominant failure of VLM policies.
- Method: a perception module predicts objects in view and goal distance each frame; the policy consumes these predictions plus observation history. Scene annotations supervise grounding per frame, so better perception improves navigation without retraining the policy.
- Evaluation flaw: official scoring rejects valid stops due to annotation defects; corrected evaluation raises success by 13 points for LightNav-0, 16 for StopWatch and 19 for ground-truth grounding. The vocabulary merges 2,408 raw labels into 540 object classes.
Code, models and data are open source.
More from Embodied
- DreamZero debuts Dream-1: zero-shot robot brain beats Opus 5.5 at 16x lower cost — gharik · 2026-10-20
- New partnership connects global robotics teams to China's humanoid hardware supply chain — MarwaEldiwiny · 2026-10-20
- Gemini turns a minimal sketch into a full working humanoid robot control sim — bousmalis · 2026-10-20
- OpenLolo: an MCP server that lets AI agents tap, swipe and type on your iPhone — Eonminion · 2026-10-20
- esp32sim-mcp: MCP server lets AI agents build and test ESP32 firmware in a local emulator — EmanueleStrazzullo · 2026-10-20
- Putting Linux on my RTX Spark laptop because I don't have a Microsoft account — pfeifferj · 2026-10-20